Upload README.md with huggingface_hub
Browse files
README.md
CHANGED
|
@@ -36,64 +36,64 @@ More details on model performance across various devices, can be found
|
|
| 36 |
|
| 37 |
| Model | Device | Chipset | Target Runtime | Inference Time (ms) | Peak Memory Range (MB) | Precision | Primary Compute Unit | Target Model
|
| 38 |
|---|---|---|---|---|---|---|---|---|
|
| 39 |
-
| TrOCRDecoder | Samsung Galaxy S23 | Snapdragon® 8 Gen 2 | TFLITE | 2.
|
| 40 |
-
| TrOCRDecoder | Samsung Galaxy S23 | Snapdragon® 8 Gen 2 | QNN | 2.
|
| 41 |
-
| TrOCRDecoder | Samsung Galaxy S23 | Snapdragon® 8 Gen 2 | ONNX | 2.
|
| 42 |
-
| TrOCRDecoder | Samsung Galaxy S24 | Snapdragon® 8 Gen 3 | TFLITE | 1.
|
| 43 |
-
| TrOCRDecoder | Samsung Galaxy S24 | Snapdragon® 8 Gen 3 | QNN | 1.
|
| 44 |
-
| TrOCRDecoder | Samsung Galaxy S24 | Snapdragon® 8 Gen 3 | ONNX | 1.
|
| 45 |
-
| TrOCRDecoder | Snapdragon 8 Elite QRD | Snapdragon® 8 Elite | TFLITE | 1.
|
| 46 |
-
| TrOCRDecoder | Snapdragon 8 Elite QRD | Snapdragon® 8 Elite | QNN | 1.
|
| 47 |
-
| TrOCRDecoder | Snapdragon 8 Elite QRD | Snapdragon® 8 Elite | ONNX | 1.
|
| 48 |
-
| TrOCRDecoder | SA7255P ADP | SA7255P | TFLITE | 11.
|
| 49 |
-
| TrOCRDecoder | SA7255P ADP | SA7255P | QNN | 11.
|
| 50 |
-
| TrOCRDecoder | SA8255 (Proxy) | SA8255P Proxy | TFLITE | 2.
|
| 51 |
-
| TrOCRDecoder | SA8255 (Proxy) | SA8255P Proxy | QNN | 2.
|
| 52 |
-
| TrOCRDecoder | SA8295P ADP | SA8295P | TFLITE |
|
| 53 |
-
| TrOCRDecoder | SA8295P ADP | SA8295P | QNN | 2.
|
| 54 |
-
| TrOCRDecoder | SA8650 (Proxy) | SA8650P Proxy | TFLITE | 2.
|
| 55 |
-
| TrOCRDecoder | SA8650 (Proxy) | SA8650P Proxy | QNN | 2.
|
| 56 |
-
| TrOCRDecoder | SA8775P ADP | SA8775P | TFLITE | 3.
|
| 57 |
-
| TrOCRDecoder | SA8775P ADP | SA8775P | QNN | 3.009 ms |
|
| 58 |
-
| TrOCRDecoder | QCS8275 (Proxy) | QCS8275 Proxy | TFLITE | 11.
|
| 59 |
-
| TrOCRDecoder | QCS8275 (Proxy) | QCS8275 Proxy | QNN | 11.
|
| 60 |
-
| TrOCRDecoder | QCS8550 (Proxy) | QCS8550 Proxy | TFLITE | 2.
|
| 61 |
-
| TrOCRDecoder | QCS8550 (Proxy) | QCS8550 Proxy | QNN | 2.
|
| 62 |
-
| TrOCRDecoder | QCS9075 (Proxy) | QCS9075 Proxy | TFLITE | 3.
|
| 63 |
-
| TrOCRDecoder | QCS9075 (Proxy) | QCS9075 Proxy | QNN | 3.009 ms |
|
| 64 |
-
| TrOCRDecoder | QCS8450 (Proxy) | QCS8450 Proxy | TFLITE | 2.
|
| 65 |
-
| TrOCRDecoder | QCS8450 (Proxy) | QCS8450 Proxy | QNN | 2.
|
| 66 |
-
| TrOCRDecoder | Snapdragon X Elite CRD | Snapdragon® X Elite | QNN | 2.
|
| 67 |
-
| TrOCRDecoder | Snapdragon X Elite CRD | Snapdragon® X Elite | ONNX | 2.
|
| 68 |
-
| TrOCREncoder | Samsung Galaxy S23 | Snapdragon® 8 Gen 2 | TFLITE | 38.
|
| 69 |
-
| TrOCREncoder | Samsung Galaxy S23 | Snapdragon® 8 Gen 2 | QNN | 36.
|
| 70 |
-
| TrOCREncoder | Samsung Galaxy S23 | Snapdragon® 8 Gen 2 | ONNX |
|
| 71 |
-
| TrOCREncoder | Samsung Galaxy S24 | Snapdragon® 8 Gen 3 | TFLITE | 29.
|
| 72 |
-
| TrOCREncoder | Samsung Galaxy S24 | Snapdragon® 8 Gen 3 | QNN |
|
| 73 |
-
| TrOCREncoder | Samsung Galaxy S24 | Snapdragon® 8 Gen 3 | ONNX | 30.
|
| 74 |
-
| TrOCREncoder | Snapdragon 8 Elite QRD | Snapdragon® 8 Elite | TFLITE | 27.
|
| 75 |
-
| TrOCREncoder | Snapdragon 8 Elite QRD | Snapdragon® 8 Elite | QNN | 26.
|
| 76 |
-
| TrOCREncoder | Snapdragon 8 Elite QRD | Snapdragon® 8 Elite | ONNX |
|
| 77 |
-
| TrOCREncoder | SA7255P ADP | SA7255P | TFLITE | 256.
|
| 78 |
-
| TrOCREncoder | SA7255P ADP | SA7255P | QNN | 248.
|
| 79 |
-
| TrOCREncoder | SA8255 (Proxy) | SA8255P Proxy | TFLITE |
|
| 80 |
-
| TrOCREncoder | SA8255 (Proxy) | SA8255P Proxy | QNN |
|
| 81 |
-
| TrOCREncoder | SA8295P ADP | SA8295P | TFLITE | 53.
|
| 82 |
-
| TrOCREncoder | SA8295P ADP | SA8295P | QNN | 50.
|
| 83 |
-
| TrOCREncoder | SA8650 (Proxy) | SA8650P Proxy | TFLITE |
|
| 84 |
-
| TrOCREncoder | SA8650 (Proxy) | SA8650P Proxy | QNN | 36.
|
| 85 |
-
| TrOCREncoder | SA8775P ADP | SA8775P | TFLITE | 47.
|
| 86 |
-
| TrOCREncoder | SA8775P ADP | SA8775P | QNN | 42.
|
| 87 |
-
| TrOCREncoder | QCS8275 (Proxy) | QCS8275 Proxy | TFLITE | 256.
|
| 88 |
-
| TrOCREncoder | QCS8275 (Proxy) | QCS8275 Proxy | QNN | 248.
|
| 89 |
-
| TrOCREncoder | QCS8550 (Proxy) | QCS8550 Proxy | TFLITE | 38.
|
| 90 |
-
| TrOCREncoder | QCS8550 (Proxy) | QCS8550 Proxy | QNN |
|
| 91 |
-
| TrOCREncoder | QCS9075 (Proxy) | QCS9075 Proxy | TFLITE | 47.
|
| 92 |
-
| TrOCREncoder | QCS9075 (Proxy) | QCS9075 Proxy | QNN | 42.
|
| 93 |
-
| TrOCREncoder | QCS8450 (Proxy) | QCS8450 Proxy | TFLITE |
|
| 94 |
-
| TrOCREncoder | QCS8450 (Proxy) | QCS8450 Proxy | QNN |
|
| 95 |
-
| TrOCREncoder | Snapdragon X Elite CRD | Snapdragon® X Elite | QNN | 34.
|
| 96 |
-
| TrOCREncoder | Snapdragon X Elite CRD | Snapdragon® X Elite | ONNX |
|
| 97 |
|
| 98 |
|
| 99 |
|
|
@@ -158,7 +158,7 @@ TrOCRDecoder
|
|
| 158 |
Device : Samsung Galaxy S23 (13)
|
| 159 |
Runtime : TFLITE
|
| 160 |
Estimated inference time (ms) : 2.1
|
| 161 |
-
Estimated peak memory usage (MB): [0,
|
| 162 |
Total # Ops : 399
|
| 163 |
Compute Unit(s) : NPU (399 ops)
|
| 164 |
|
|
@@ -167,7 +167,7 @@ TrOCREncoder
|
|
| 167 |
Device : Samsung Galaxy S23 (13)
|
| 168 |
Runtime : TFLITE
|
| 169 |
Estimated inference time (ms) : 38.9
|
| 170 |
-
Estimated peak memory usage (MB): [7,
|
| 171 |
Total # Ops : 591
|
| 172 |
Compute Unit(s) : NPU (591 ops)
|
| 173 |
```
|
|
@@ -191,43 +191,26 @@ import qai_hub as hub
|
|
| 191 |
from qai_hub_models.models.trocr import Model
|
| 192 |
|
| 193 |
# Load the model
|
| 194 |
-
|
| 195 |
-
decoder_model = model.decoder
|
| 196 |
-
encoder_model = model.encoder
|
| 197 |
|
| 198 |
# Device
|
| 199 |
-
device = hub.Device("Samsung Galaxy
|
| 200 |
|
| 201 |
# Trace model
|
| 202 |
-
|
| 203 |
-
|
| 204 |
|
| 205 |
-
|
| 206 |
|
| 207 |
# Compile model on a specific device
|
| 208 |
-
|
| 209 |
-
model=
|
| 210 |
device=device,
|
| 211 |
-
input_specs=
|
| 212 |
)
|
| 213 |
|
| 214 |
# Get target model to run on-device
|
| 215 |
-
|
| 216 |
-
# Trace model
|
| 217 |
-
encoder_input_shape = encoder_model.get_input_spec()
|
| 218 |
-
encoder_sample_inputs = encoder_model.sample_inputs()
|
| 219 |
-
|
| 220 |
-
traced_encoder_model = torch.jit.trace(encoder_model, [torch.tensor(data[0]) for _, data in encoder_sample_inputs.items()])
|
| 221 |
-
|
| 222 |
-
# Compile model on a specific device
|
| 223 |
-
encoder_compile_job = hub.submit_compile_job(
|
| 224 |
-
model=traced_encoder_model ,
|
| 225 |
-
device=device,
|
| 226 |
-
input_specs=encoder_model.get_input_spec(),
|
| 227 |
-
)
|
| 228 |
-
|
| 229 |
-
# Get target model to run on-device
|
| 230 |
-
encoder_target_model = encoder_compile_job.get_target_model()
|
| 231 |
|
| 232 |
```
|
| 233 |
|
|
@@ -239,15 +222,11 @@ After compiling models from step 1. Models can be profiled model on-device using
|
|
| 239 |
provisioned in the cloud. Once the job is submitted, you can navigate to a
|
| 240 |
provided job URL to view a variety of on-device performance metrics.
|
| 241 |
```python
|
| 242 |
-
|
| 243 |
-
model=
|
| 244 |
-
device=device,
|
| 245 |
-
)
|
| 246 |
-
encoder_profile_job = hub.submit_profile_job(
|
| 247 |
-
model=encoder_target_model,
|
| 248 |
device=device,
|
| 249 |
)
|
| 250 |
-
|
| 251 |
```
|
| 252 |
|
| 253 |
Step 3: **Verify on-device accuracy**
|
|
@@ -255,20 +234,13 @@ Step 3: **Verify on-device accuracy**
|
|
| 255 |
To verify the accuracy of the model on-device, you can run on-device inference
|
| 256 |
on sample input data on the same cloud hosted device.
|
| 257 |
```python
|
| 258 |
-
|
| 259 |
-
|
| 260 |
-
model=
|
| 261 |
-
device=device,
|
| 262 |
-
inputs=decoder_input_data,
|
| 263 |
-
)
|
| 264 |
-
decoder_inference_job.download_output_data()
|
| 265 |
-
encoder_input_data = encoder_model.sample_inputs()
|
| 266 |
-
encoder_inference_job = hub.submit_inference_job(
|
| 267 |
-
model=encoder_target_model,
|
| 268 |
device=device,
|
| 269 |
-
inputs=
|
| 270 |
)
|
| 271 |
-
|
| 272 |
|
| 273 |
```
|
| 274 |
With the output of the model, you can compute like PSNR, relative errors or
|
|
|
|
| 36 |
|
| 37 |
| Model | Device | Chipset | Target Runtime | Inference Time (ms) | Peak Memory Range (MB) | Precision | Primary Compute Unit | Target Model
|
| 38 |
|---|---|---|---|---|---|---|---|---|
|
| 39 |
+
| TrOCRDecoder | Samsung Galaxy S23 | Snapdragon® 8 Gen 2 | TFLITE | 2.061 ms | 0 - 27 MB | FP16 | NPU | [TrOCR.tflite](https://huggingface.co/qualcomm/TrOCR/blob/main/TrOCRDecoder.tflite) |
|
| 40 |
+
| TrOCRDecoder | Samsung Galaxy S23 | Snapdragon® 8 Gen 2 | QNN | 2.083 ms | 2 - 4 MB | FP16 | NPU | [TrOCR.so](https://huggingface.co/qualcomm/TrOCR/blob/main/TrOCRDecoder.so) |
|
| 41 |
+
| TrOCRDecoder | Samsung Galaxy S23 | Snapdragon® 8 Gen 2 | ONNX | 2.55 ms | 0 - 128 MB | FP16 | NPU | [TrOCR.onnx](https://huggingface.co/qualcomm/TrOCR/blob/main/TrOCRDecoder.onnx) |
|
| 42 |
+
| TrOCRDecoder | Samsung Galaxy S24 | Snapdragon® 8 Gen 3 | TFLITE | 1.512 ms | 0 - 135 MB | FP16 | NPU | [TrOCR.tflite](https://huggingface.co/qualcomm/TrOCR/blob/main/TrOCRDecoder.tflite) |
|
| 43 |
+
| TrOCRDecoder | Samsung Galaxy S24 | Snapdragon® 8 Gen 3 | QNN | 1.509 ms | 0 - 18 MB | FP16 | NPU | [TrOCR.so](https://huggingface.co/qualcomm/TrOCR/blob/main/TrOCRDecoder.so) |
|
| 44 |
+
| TrOCRDecoder | Samsung Galaxy S24 | Snapdragon® 8 Gen 3 | ONNX | 1.766 ms | 0 - 135 MB | FP16 | NPU | [TrOCR.onnx](https://huggingface.co/qualcomm/TrOCR/blob/main/TrOCRDecoder.onnx) |
|
| 45 |
+
| TrOCRDecoder | Snapdragon 8 Elite QRD | Snapdragon® 8 Elite | TFLITE | 1.228 ms | 0 - 61 MB | FP16 | NPU | [TrOCR.tflite](https://huggingface.co/qualcomm/TrOCR/blob/main/TrOCRDecoder.tflite) |
|
| 46 |
+
| TrOCRDecoder | Snapdragon 8 Elite QRD | Snapdragon® 8 Elite | QNN | 1.404 ms | 3 - 65 MB | FP16 | NPU | Use Export Script |
|
| 47 |
+
| TrOCRDecoder | Snapdragon 8 Elite QRD | Snapdragon® 8 Elite | ONNX | 1.678 ms | 3 - 152 MB | FP16 | NPU | [TrOCR.onnx](https://huggingface.co/qualcomm/TrOCR/blob/main/TrOCRDecoder.onnx) |
|
| 48 |
+
| TrOCRDecoder | SA7255P ADP | SA7255P | TFLITE | 11.747 ms | 0 - 60 MB | FP16 | NPU | [TrOCR.tflite](https://huggingface.co/qualcomm/TrOCR/blob/main/TrOCRDecoder.tflite) |
|
| 49 |
+
| TrOCRDecoder | SA7255P ADP | SA7255P | QNN | 11.72 ms | 7 - 16 MB | FP16 | NPU | Use Export Script |
|
| 50 |
+
| TrOCRDecoder | SA8255 (Proxy) | SA8255P Proxy | TFLITE | 2.083 ms | 0 - 30 MB | FP16 | NPU | [TrOCR.tflite](https://huggingface.co/qualcomm/TrOCR/blob/main/TrOCRDecoder.tflite) |
|
| 51 |
+
| TrOCRDecoder | SA8255 (Proxy) | SA8255P Proxy | QNN | 2.041 ms | 1 - 3 MB | FP16 | NPU | Use Export Script |
|
| 52 |
+
| TrOCRDecoder | SA8295P ADP | SA8295P | TFLITE | 3.033 ms | 0 - 52 MB | FP16 | NPU | [TrOCR.tflite](https://huggingface.co/qualcomm/TrOCR/blob/main/TrOCRDecoder.tflite) |
|
| 53 |
+
| TrOCRDecoder | SA8295P ADP | SA8295P | QNN | 2.893 ms | 7 - 25 MB | FP16 | NPU | Use Export Script |
|
| 54 |
+
| TrOCRDecoder | SA8650 (Proxy) | SA8650P Proxy | TFLITE | 2.1 ms | 0 - 30 MB | FP16 | NPU | [TrOCR.tflite](https://huggingface.co/qualcomm/TrOCR/blob/main/TrOCRDecoder.tflite) |
|
| 55 |
+
| TrOCRDecoder | SA8650 (Proxy) | SA8650P Proxy | QNN | 2.015 ms | 2 - 5 MB | FP16 | NPU | Use Export Script |
|
| 56 |
+
| TrOCRDecoder | SA8775P ADP | SA8775P | TFLITE | 3.149 ms | 0 - 59 MB | FP16 | NPU | [TrOCR.tflite](https://huggingface.co/qualcomm/TrOCR/blob/main/TrOCRDecoder.tflite) |
|
| 57 |
+
| TrOCRDecoder | SA8775P ADP | SA8775P | QNN | 3.009 ms | 7 - 17 MB | FP16 | NPU | Use Export Script |
|
| 58 |
+
| TrOCRDecoder | QCS8275 (Proxy) | QCS8275 Proxy | TFLITE | 11.747 ms | 0 - 60 MB | FP16 | NPU | [TrOCR.tflite](https://huggingface.co/qualcomm/TrOCR/blob/main/TrOCRDecoder.tflite) |
|
| 59 |
+
| TrOCRDecoder | QCS8275 (Proxy) | QCS8275 Proxy | QNN | 11.72 ms | 7 - 16 MB | FP16 | NPU | Use Export Script |
|
| 60 |
+
| TrOCRDecoder | QCS8550 (Proxy) | QCS8550 Proxy | TFLITE | 2.066 ms | 0 - 31 MB | FP16 | NPU | [TrOCR.tflite](https://huggingface.co/qualcomm/TrOCR/blob/main/TrOCRDecoder.tflite) |
|
| 61 |
+
| TrOCRDecoder | QCS8550 (Proxy) | QCS8550 Proxy | QNN | 2.03 ms | 2 - 5 MB | FP16 | NPU | Use Export Script |
|
| 62 |
+
| TrOCRDecoder | QCS9075 (Proxy) | QCS9075 Proxy | TFLITE | 3.149 ms | 0 - 59 MB | FP16 | NPU | [TrOCR.tflite](https://huggingface.co/qualcomm/TrOCR/blob/main/TrOCRDecoder.tflite) |
|
| 63 |
+
| TrOCRDecoder | QCS9075 (Proxy) | QCS9075 Proxy | QNN | 3.009 ms | 7 - 17 MB | FP16 | NPU | Use Export Script |
|
| 64 |
+
| TrOCRDecoder | QCS8450 (Proxy) | QCS8450 Proxy | TFLITE | 2.885 ms | 0 - 122 MB | FP16 | NPU | [TrOCR.tflite](https://huggingface.co/qualcomm/TrOCR/blob/main/TrOCRDecoder.tflite) |
|
| 65 |
+
| TrOCRDecoder | QCS8450 (Proxy) | QCS8450 Proxy | QNN | 2.985 ms | 7 - 133 MB | FP16 | NPU | Use Export Script |
|
| 66 |
+
| TrOCRDecoder | Snapdragon X Elite CRD | Snapdragon® X Elite | QNN | 2.234 ms | 7 - 7 MB | FP16 | NPU | Use Export Script |
|
| 67 |
+
| TrOCRDecoder | Snapdragon X Elite CRD | Snapdragon® X Elite | ONNX | 2.322 ms | 68 - 68 MB | FP16 | NPU | [TrOCR.onnx](https://huggingface.co/qualcomm/TrOCR/blob/main/TrOCRDecoder.onnx) |
|
| 68 |
+
| TrOCREncoder | Samsung Galaxy S23 | Snapdragon® 8 Gen 2 | TFLITE | 38.875 ms | 7 - 29 MB | FP16 | NPU | [TrOCR.tflite](https://huggingface.co/qualcomm/TrOCR/blob/main/TrOCREncoder.tflite) |
|
| 69 |
+
| TrOCREncoder | Samsung Galaxy S23 | Snapdragon® 8 Gen 2 | QNN | 36.998 ms | 0 - 2 MB | FP16 | NPU | [TrOCR.so](https://huggingface.co/qualcomm/TrOCR/blob/main/TrOCREncoder.so) |
|
| 70 |
+
| TrOCREncoder | Samsung Galaxy S23 | Snapdragon® 8 Gen 2 | ONNX | 36.962 ms | 7 - 134 MB | FP16 | NPU | [TrOCR.onnx](https://huggingface.co/qualcomm/TrOCR/blob/main/TrOCREncoder.onnx) |
|
| 71 |
+
| TrOCREncoder | Samsung Galaxy S24 | Snapdragon® 8 Gen 3 | TFLITE | 29.99 ms | 6 - 174 MB | FP16 | NPU | [TrOCR.tflite](https://huggingface.co/qualcomm/TrOCR/blob/main/TrOCREncoder.tflite) |
|
| 72 |
+
| TrOCREncoder | Samsung Galaxy S24 | Snapdragon® 8 Gen 3 | QNN | 30.188 ms | 2 - 20 MB | FP16 | NPU | [TrOCR.so](https://huggingface.co/qualcomm/TrOCR/blob/main/TrOCREncoder.so) |
|
| 73 |
+
| TrOCREncoder | Samsung Galaxy S24 | Snapdragon® 8 Gen 3 | ONNX | 30.736 ms | 14 - 165 MB | FP16 | NPU | [TrOCR.onnx](https://huggingface.co/qualcomm/TrOCR/blob/main/TrOCREncoder.onnx) |
|
| 74 |
+
| TrOCREncoder | Snapdragon 8 Elite QRD | Snapdragon® 8 Elite | TFLITE | 27.772 ms | 5 - 168 MB | FP16 | NPU | [TrOCR.tflite](https://huggingface.co/qualcomm/TrOCR/blob/main/TrOCREncoder.tflite) |
|
| 75 |
+
| TrOCREncoder | Snapdragon 8 Elite QRD | Snapdragon® 8 Elite | QNN | 26.54 ms | 2 - 166 MB | FP16 | NPU | Use Export Script |
|
| 76 |
+
| TrOCREncoder | Snapdragon 8 Elite QRD | Snapdragon® 8 Elite | ONNX | 28.836 ms | 14 - 172 MB | FP16 | NPU | [TrOCR.onnx](https://huggingface.co/qualcomm/TrOCR/blob/main/TrOCREncoder.onnx) |
|
| 77 |
+
| TrOCREncoder | SA7255P ADP | SA7255P | TFLITE | 256.881 ms | 7 - 170 MB | FP16 | NPU | [TrOCR.tflite](https://huggingface.co/qualcomm/TrOCR/blob/main/TrOCREncoder.tflite) |
|
| 78 |
+
| TrOCREncoder | SA7255P ADP | SA7255P | QNN | 248.315 ms | 2 - 11 MB | FP16 | NPU | Use Export Script |
|
| 79 |
+
| TrOCREncoder | SA8255 (Proxy) | SA8255P Proxy | TFLITE | 38.924 ms | 8 - 32 MB | FP16 | NPU | [TrOCR.tflite](https://huggingface.co/qualcomm/TrOCR/blob/main/TrOCREncoder.tflite) |
|
| 80 |
+
| TrOCREncoder | SA8255 (Proxy) | SA8255P Proxy | QNN | 36.621 ms | 2 - 5 MB | FP16 | NPU | Use Export Script |
|
| 81 |
+
| TrOCREncoder | SA8295P ADP | SA8295P | TFLITE | 53.159 ms | 7 - 169 MB | FP16 | NPU | [TrOCR.tflite](https://huggingface.co/qualcomm/TrOCR/blob/main/TrOCREncoder.tflite) |
|
| 82 |
+
| TrOCREncoder | SA8295P ADP | SA8295P | QNN | 50.227 ms | 2 - 18 MB | FP16 | NPU | Use Export Script |
|
| 83 |
+
| TrOCREncoder | SA8650 (Proxy) | SA8650P Proxy | TFLITE | 38.668 ms | 7 - 36 MB | FP16 | NPU | [TrOCR.tflite](https://huggingface.co/qualcomm/TrOCR/blob/main/TrOCREncoder.tflite) |
|
| 84 |
+
| TrOCREncoder | SA8650 (Proxy) | SA8650P Proxy | QNN | 36.86 ms | 2 - 4 MB | FP16 | NPU | Use Export Script |
|
| 85 |
+
| TrOCREncoder | SA8775P ADP | SA8775P | TFLITE | 47.569 ms | 7 - 169 MB | FP16 | NPU | [TrOCR.tflite](https://huggingface.co/qualcomm/TrOCR/blob/main/TrOCREncoder.tflite) |
|
| 86 |
+
| TrOCREncoder | SA8775P ADP | SA8775P | QNN | 42.56 ms | 2 - 11 MB | FP16 | NPU | Use Export Script |
|
| 87 |
+
| TrOCREncoder | QCS8275 (Proxy) | QCS8275 Proxy | TFLITE | 256.881 ms | 7 - 170 MB | FP16 | NPU | [TrOCR.tflite](https://huggingface.co/qualcomm/TrOCR/blob/main/TrOCREncoder.tflite) |
|
| 88 |
+
| TrOCREncoder | QCS8275 (Proxy) | QCS8275 Proxy | QNN | 248.315 ms | 2 - 11 MB | FP16 | NPU | Use Export Script |
|
| 89 |
+
| TrOCREncoder | QCS8550 (Proxy) | QCS8550 Proxy | TFLITE | 38.449 ms | 7 - 32 MB | FP16 | NPU | [TrOCR.tflite](https://huggingface.co/qualcomm/TrOCR/blob/main/TrOCREncoder.tflite) |
|
| 90 |
+
| TrOCREncoder | QCS8550 (Proxy) | QCS8550 Proxy | QNN | 37.026 ms | 2 - 5 MB | FP16 | NPU | Use Export Script |
|
| 91 |
+
| TrOCREncoder | QCS9075 (Proxy) | QCS9075 Proxy | TFLITE | 47.569 ms | 7 - 169 MB | FP16 | NPU | [TrOCR.tflite](https://huggingface.co/qualcomm/TrOCR/blob/main/TrOCREncoder.tflite) |
|
| 92 |
+
| TrOCREncoder | QCS9075 (Proxy) | QCS9075 Proxy | QNN | 42.56 ms | 2 - 11 MB | FP16 | NPU | Use Export Script |
|
| 93 |
+
| TrOCREncoder | QCS8450 (Proxy) | QCS8450 Proxy | TFLITE | 48.787 ms | 7 - 171 MB | FP16 | NPU | [TrOCR.tflite](https://huggingface.co/qualcomm/TrOCR/blob/main/TrOCREncoder.tflite) |
|
| 94 |
+
| TrOCREncoder | QCS8450 (Proxy) | QCS8450 Proxy | QNN | 50.666 ms | 2 - 168 MB | FP16 | NPU | Use Export Script |
|
| 95 |
+
| TrOCREncoder | Snapdragon X Elite CRD | Snapdragon® X Elite | QNN | 34.302 ms | 2 - 2 MB | FP16 | NPU | Use Export Script |
|
| 96 |
+
| TrOCREncoder | Snapdragon X Elite CRD | Snapdragon® X Elite | ONNX | 36.636 ms | 50 - 50 MB | FP16 | NPU | [TrOCR.onnx](https://huggingface.co/qualcomm/TrOCR/blob/main/TrOCREncoder.onnx) |
|
| 97 |
|
| 98 |
|
| 99 |
|
|
|
|
| 158 |
Device : Samsung Galaxy S23 (13)
|
| 159 |
Runtime : TFLITE
|
| 160 |
Estimated inference time (ms) : 2.1
|
| 161 |
+
Estimated peak memory usage (MB): [0, 27]
|
| 162 |
Total # Ops : 399
|
| 163 |
Compute Unit(s) : NPU (399 ops)
|
| 164 |
|
|
|
|
| 167 |
Device : Samsung Galaxy S23 (13)
|
| 168 |
Runtime : TFLITE
|
| 169 |
Estimated inference time (ms) : 38.9
|
| 170 |
+
Estimated peak memory usage (MB): [7, 29]
|
| 171 |
Total # Ops : 591
|
| 172 |
Compute Unit(s) : NPU (591 ops)
|
| 173 |
```
|
|
|
|
| 191 |
from qai_hub_models.models.trocr import Model
|
| 192 |
|
| 193 |
# Load the model
|
| 194 |
+
torch_model = Model.from_pretrained()
|
|
|
|
|
|
|
| 195 |
|
| 196 |
# Device
|
| 197 |
+
device = hub.Device("Samsung Galaxy S24")
|
| 198 |
|
| 199 |
# Trace model
|
| 200 |
+
input_shape = torch_model.get_input_spec()
|
| 201 |
+
sample_inputs = torch_model.sample_inputs()
|
| 202 |
|
| 203 |
+
pt_model = torch.jit.trace(torch_model, [torch.tensor(data[0]) for _, data in sample_inputs.items()])
|
| 204 |
|
| 205 |
# Compile model on a specific device
|
| 206 |
+
compile_job = hub.submit_compile_job(
|
| 207 |
+
model=pt_model,
|
| 208 |
device=device,
|
| 209 |
+
input_specs=torch_model.get_input_spec(),
|
| 210 |
)
|
| 211 |
|
| 212 |
# Get target model to run on-device
|
| 213 |
+
target_model = compile_job.get_target_model()
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 214 |
|
| 215 |
```
|
| 216 |
|
|
|
|
| 222 |
provisioned in the cloud. Once the job is submitted, you can navigate to a
|
| 223 |
provided job URL to view a variety of on-device performance metrics.
|
| 224 |
```python
|
| 225 |
+
profile_job = hub.submit_profile_job(
|
| 226 |
+
model=target_model,
|
|
|
|
|
|
|
|
|
|
|
|
|
| 227 |
device=device,
|
| 228 |
)
|
| 229 |
+
|
| 230 |
```
|
| 231 |
|
| 232 |
Step 3: **Verify on-device accuracy**
|
|
|
|
| 234 |
To verify the accuracy of the model on-device, you can run on-device inference
|
| 235 |
on sample input data on the same cloud hosted device.
|
| 236 |
```python
|
| 237 |
+
input_data = torch_model.sample_inputs()
|
| 238 |
+
inference_job = hub.submit_inference_job(
|
| 239 |
+
model=target_model,
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 240 |
device=device,
|
| 241 |
+
inputs=input_data,
|
| 242 |
)
|
| 243 |
+
on_device_output = inference_job.download_output_data()
|
| 244 |
|
| 245 |
```
|
| 246 |
With the output of the model, you can compute like PSNR, relative errors or
|